LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images.

Jing Zhang*, Irving Fang*, Hao Wu, Akshat Kaushik, Alice Rodriguez, Hanwen Zhao, Juexiao Zhang, Zhuo Zheng, Radu Iovita, Chen Feng

Project Website

Please visit our project website for more information, including an interactive demo with real artifacts.

Environment Setup

The project was developed on Python 3.11.5 and PyTorch 2.1.1 with CUDA 11.8.0 binaries. While you can refer to requirements.txt for more details, having PyTorch installed should be mostly enough for this project.

We utilized PyTorch 2.1.1 to access some of its exclusive features such as torch.compile() to accelerate training as much as we could. However, most (if not all) of these techniques should not affect the inference accuracy, so you should be able to perfectly replicate our results without a matching PyTorch version.

Dataset

Please visit our Hugging Face repo to access the dataset.

Please refer to transfer_learning/data_utils/data_tribology.py for how to process them.

1. we use integers to label stone that has been worked again certain material as follows:

   Material	Integer
   ANTLER	0
   BEECHWOOD	1
   BEFOREUSE	2
   BONE	3
   IVORY	4
   SPRUCEWOOD	5
   BARLEY	6
   FERN	7
   HORSETAIL	8

   Here, "BEFOREUSE" refers to a state where the stone is not polished with any material at all.

2. In the dataset, $256, 512, 865$ refers to the resolution of the images.

   a. The images were originally taken at the resolution of $865 \times 865$. This corresponds to the 1 granularity in the paper.

   b. $512$ corresponds to the 6 granularity in the paper.

   c. $256$ corresponds to the 24 granularity in the paper.

Fully-Supervised Image Classification

To reproduce the results in the Fully-Supervised Image Classification section of the paper, please refer to the transfer_learning folder.

1. To train a specific deep learning model, please run the following command.

   python dl_supervised_pipeline.py \
   --resolution "$RESOLUTION" \
   --magnification "$MAGNIFICATION" \
   --modality "$MODALITY" \
   --model "$MODEL" \
   --pretrained $PRETRAINED \
   --frozen $FROZEN \
   --vote $VOTE \
   --epochs $EPOCHS \
   --batch_size $BATCH_SIZE \
   --start_lr $START_LR \
   --seed $SEED
   

   You can take a look at any given script in transfer_learning /launch_scripts/ for reference.

2. If you want to replicate the SVM-based model, please instead run the svm_pipeline.py

3. There are some helper functions located in transfer_learning/experiments/collect_results.py to gather the inference results in a more presentable and readable format. Although ironically the helper script itself is not super tidy.

Few-Shot Image Classification

To reproduce the results in the Few-Shot Image Classification section of the paper, please refer to the fewshot_learning folder.

Citation

If you find our work helpful in your research, please consider citing the following:

@InProceedings{Zhang_2024_CVPR,
    author    = {Zhang, Jing and Fang, Irving and Wu, Hao and Kaushik, Akshat and Rodriguez, Alice and Zhao, Hanwen and Zhang, Juexiao and Zheng, Zhuo and Iovita, Radu and Feng, Chen},
    title     = {LUWA Dataset: Learning Lithic Use-Wear Analysis on Microscopic Images},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2024},
    pages     = {22563-22573}
}